Abliterated models: what they are and where to run them
Abliteration removes the refusal direction from a model’s weights. The abliterated Llama, Qwen and NeuralDaredevil models we serve, with prices, and how to call them through an OpenAI-compatible API.
Chat with an abliterated model →
Short answer
An abliterated model is an open model whose "refusal direction" has been found and removed from its weights, so it stops saying no without being retrained. You can run one locally if you have the hardware, or call one here with no setup.
Abliterated models we serve
nano-llama-3.3-70b-instruct-abliteratedandfeatherless-llama-3.3-70b-instruct-abliterated: Llama 3.3 70B.featherless-qwen2.5-72b-instruct-abliterated: Qwen 2.5 72B.nano-qwen2.5-32b-instruct-abliteratedandfeatherless-qwen2.5-32b-instruct-abliterated: Qwen 2.5 32B.featherless-neuraldaredevil-8b-abliterated: a small, fast 8B model.- Live prices for each are in the model catalog.
How it works, in one paragraph
Researchers found that refusals in many chat models are driven by a single direction in the model’s internal activations. Projecting that direction out of the weights ("orthogonalizing" them) removes most refusals while keeping the model’s other abilities. Full explanation: what is an abliterated model.
FAQ
Are abliterated models legal to use?
Yes. They are open-weight models with a modification; what you do with the output is still subject to the law and our acceptable-use rules.
Do I need to be 18?
Yes. Our terms require users to be at least 18.
Can I use my own characters?
Yes. Paste a character description into the chat, or use our API as the backend for SillyTavern or JanitorAI and keep your existing character cards. See SillyTavern setup and JanitorAI proxy setup.